U.S. patent application number 16/875041 was filed with the patent office on 2020-11-19 for systems and methods for predicting pain level.
This patent application is currently assigned to The Florida International University Board of Trustees. The applicant listed for this patent is Deya Banisakher, Mark Finlayson, Naphtali Rishe. Invention is credited to Deya Banisakher, Mark Finlayson, Naphtali Rishe.
Application Number | 20200364566 16/875041 |
Document ID | / |
Family ID | 1000004868079 |
Filed Date | 2020-11-19 |
United States Patent
Application |
20200364566 |
Kind Code |
A1 |
Banisakher; Deya ; et
al. |
November 19, 2020 |
SYSTEMS AND METHODS FOR PREDICTING PAIN LEVEL
Abstract
Devices and methods for learning and/or predicting the
self-reported pain improvement levels of osteoarthritis (OA)
patients are provided. A device or apparatus can include a
processor and a machine-readable medium in operable communication
with the processor and having stored thereon an algorithm and a
unique set of features. The algorithm and set of features can
enable building one or more models that learn the self-reported
pain improvement levels of OA patients.
Inventors: |
Banisakher; Deya; (Miami,
FL) ; Rishe; Naphtali; (Miami Beach, FL) ;
Finlayson; Mark; (North Bay Village, FL) |
|
Applicant: |
Name |
City |
State |
Country |
Type |
Banisakher; Deya
Rishe; Naphtali
Finlayson; Mark |
Miami
Miami Beach
North Bay Village |
FL
FL
FL |
US
US
US |
|
|
Assignee: |
The Florida International
University Board of Trustees
Miami
FL
|
Family ID: |
1000004868079 |
Appl. No.: |
16/875041 |
Filed: |
May 15, 2020 |
Related U.S. Patent Documents
|
|
|
|
|
|
Application
Number |
Filing Date |
Patent Number |
|
|
62848179 |
May 15, 2019 |
|
|
|
Current U.S.
Class: |
1/1 |
Current CPC
Class: |
G06N 3/08 20130101; G16H
50/20 20180101; G16H 10/60 20180101; G06N 3/0445 20130101; G16H
70/60 20180101 |
International
Class: |
G06N 3/08 20060101
G06N003/08; G06N 3/04 20060101 G06N003/04; G16H 50/20 20060101
G16H050/20; G16H 70/60 20060101 G16H070/60 |
Claims
1. A system for predicting a pain level of an osteoarthritis (OA)
patient, the system comprising: a processor; and a machine-readable
medium in operable communication with the processor and having
instructions stored thereon that, when executed by the processor,
perform the following steps: developing a set of classifiers, the
set of classifiers comprising three classifiers corresponding to a
first category, a second category, and a third category,
respectively; training the set of classifiers; testing the set of
classifiers; and using the set of classifiers to predict the pain
level of the OA patient at a future visit intended to assess the
pain level, the first category being that pain has improved for the
OA patient since a previous visit, the second category being that
pain has remained unchanged for the OA patient since the previous
visit, and the third category being that pain has worsened for the
OA patient since the previous visit.
2. The system according to claim 1, the developing, training,
testing, and using of the set of classifiers comprising using a
machine learning (ML) technique.
3. The system according to claim 2, the ML technique being an
eminent support vector machine (SVM), a random decision forest
(RDF), a backpropagation neural network, or a recurrent neural
network (RNN).
4. The system according to claim 2, the ML technique being an
RNN.
5. The system according to claim 1, the training of the set of
classifiers comprising training the set of classifiers using a
dataset with known values.
6. The system according to claim 1, the testing of the set of
classifiers comprising testing the set of classifiers using a
dataset with known values.
7. The system according to claim 1, the training of the set of
classifiers comprising training the set of classifiers using a
dataset with known values, and the testing of the set of
classifiers comprising testing the set of classifiers using the
dataset with known values, and the dataset being broken into a
first sub-dataset to be used for the training of the set of
classifiers and a second sub-dataset to be used for the testing of
the set of classifiers.
8. The system according to claim 1, the developing of the set of
classifiers comprising feature selection, and the training of the
set of classifiers comprising normalization of data obtained from a
dataset with known values used to train the set of classifiers.
9. The system according to claim 1, the using of the set of
classifiers to predict the pain level of the OA patient comprising
using the set of classifiers to predict the pain level of the OA
patient at an Nth visit based on features of the set of classifiers
reported on all visits up to an (N-1)th visit.
10. The system according to claim 1, the training of the set of
classifiers comprising training the set of classifiers using a
dataset with known values, and the testing of the set of
classifiers comprising testing the set of classifiers using the
dataset with known values, and the dataset with known values being
the Osteoarthritis Initiative (OAI) dataset.
11. A method for predicting a pain level of an osteoarthritis (OA)
patient, the method comprising: developing, by a processor, a set
of classifiers, the set of classifiers comprising three classifiers
corresponding to a first category, a second category, and a third
category, respectively; training, by the processor, the set of
classifiers; testing, by the processor, the set of classifiers; and
using, by the processor, the set of classifiers to predict the pain
level of the OA patient at a future visit intended to assess the
pain level, the first category being that pain has improved for the
OA patient since a previous visit, the second category being that
pain has remained unchanged for the OA patient since the previous
visit, and the third category being that pain has worsened for the
OA patient since the previous visit.
12. The method according to claim 11, the developing, training,
testing, and using of the set of classifiers comprising using a
machine learning (ML) technique.
13. The method according to claim 12, the ML technique being an
eminent support vector machine (SVM), a random decision forest
(RDF), a backpropagation neural network, or a recurrent neural
network (RNN).
14. The method according to claim 12, the ML technique being an
RNN.
15. The method according to claim 11, the training of the set of
classifiers comprising training the set of classifiers using a
dataset with known values, and the testing of the set of
classifiers comprising testing the set of classifiers using the
dataset with known values.
16. The method according to claim 15, the dataset being broken into
a first sub-dataset to be used for the training of the set of
classifiers and a second sub-dataset to be used for the testing of
the set of classifiers.
17. The method according to claim 11, the developing of the set of
classifiers comprising feature selection, and the training of the
set of classifiers comprising normalization of data obtained from a
dataset with known values used to train the set of classifiers.
18. The method according to claim 11, the using of the set of
classifiers to predict the pain level of the OA patient comprising
using the set of classifiers to predict the pain level of the OA
patient at an Nth visit based on features of the set of classifiers
reported on all visits up to an (N-1)th visit.
19. The method according to claim 11, the training of the set of
classifiers comprising training the set of classifiers using a
dataset with known values, and the testing of the set of
classifiers comprising testing the set of classifiers using the
dataset with known values, and the dataset with known values being
the Osteoarthritis Initiative (OAI) dataset.
20. A system for predicting a pain level of an osteoarthritis (OA)
patient, the system comprising: a processor; and a machine-readable
medium in operable communication with the processor and having
instructions stored thereon that, when executed by the processor,
perform the following steps: developing a set of classifiers, the
set of classifiers comprising three classifiers corresponding to a
first category, a second category, and a third category,
respectively; training the set of classifiers; testing the set of
classifiers; and using the set of classifiers to predict the pain
level of the OA patient at a future visit intended to assess the
pain level, the first category being that pain has improved for the
OA patient since a previous visit, the second category being that
pain has remained unchanged for the OA patient since the previous
visit, and the third category being that pain has worsened for the
OA patient since the previous visit, the developing, training,
testing, and using of the set of classifiers comprising using a
machine learning (ML) technique, the ML technique being an RNN, the
training of the set of classifiers comprising training the set of
classifiers using a dataset with known values, the testing of the
set of classifiers comprising testing the set of classifiers using
the dataset with known values, the dataset being broken into a
first sub-dataset to be used for the training of the set of
classifiers and a second sub-dataset to be used for the testing of
the set of classifiers, the developing of the set of classifiers
comprising feature selection, the training of the set of
classifiers comprising normalization of data obtained from the
dataset, and the using of the set of classifiers to predict the
pain level of the OA patient comprising using the set of
classifiers to predict the pain level of the OA patient at an Nth
visit based on features of the set of classifiers reported on all
visits up to an (N-1)th visit.
Description
CROSS-REFERENCE TO A RELATED APPLICATION
[0001] This application claims the benefit of U.S. Provisional
Application Ser. No. 62/848,179, filed May 15, 2019, which is
hereby incorporated by reference herein in its entirety, including
any FIGURES, tables, and drawings.
BACKGROUND
[0002] Knee osteoarthritis (OA) is the most common joint illness in
adults around the world. Previous research has demonstrated that
the early analysis and treatment of knee OA could counteract
development of symptoms. Thus, clinicians are faced with the
challenge of recognizing patients who are at high risk of
radiographic and symptomatic knee OA and projecting their treatment
outcomes in an opportune and proper way.
[0003] The National Institute of Health (NIH) describes some of the
common features of people at high risk for OA generally: females
over 45 years of age; overweight people; and people "with jobs that
stress particular joints". To assess the connection between those
features and knee OA specifically, a few strategies have been
proposed in the past. Screening surveys for symptomatic knee OA
have been used in view of patients' self-reported side effects.
Nonetheless, such screening methods demonstrate low specificity,
and cannot predict radiographic knee OA without associated
pain.
BRIEF SUMMARY
[0004] Embodiments of the subject invention provide novel and
advantageous devices and methods for learning and/or predicting the
(self-reported) pain improvement levels of osteoarthritis (OA)
patients (e.g., knee OA patients). A device or apparatus can
include a processor and a machine-readable medium (e.g., a
(non-transitory) computer-readable medium) in operable
communication with the processor and having stored thereon an
algorithm and a unique set of features (see, e.g., Table 2 herein).
The algorithm and/or set of features can be embodied as a set of
instructions stored on the machine-readable medium that, when
executed by the processor, perform steps (including steps of the
algorithm). The algorithm and set of features can enable building
one or more models that learn the (self-reported) pain improvement
levels of OA patients (e.g., knee OA patients).
[0005] In an embodiment, a system for predicting a pain level of an
OA patient can comprise: a processor; and a machine-readable medium
in operable communication with the processor and having
instructions stored thereon that, when executed by the processor,
perform the following steps: developing a set of classifiers, the
set of classifiers comprising three classifiers corresponding to a
first category, a second category, and a third category,
respectively; training the set of classifiers; testing the set of
classifiers; and using the set of classifiers to predict the pain
level of the OA patient at a future visit intended to assess the
pain level. The first category can be that pain has improved for
the OA patient since a previous visit; the second category can be
that pain has remained unchanged for the OA patient since the
previous visit; and the third category can be that pain has
worsened for the OA patient since the previous visit. The
developing, training, testing, and using of the set of classifiers
can comprise using a machine learning (ML) technique, such as an
eminent support vector machine (SVM), a random decision forest
(RDF), a backpropagation neural network, or a recurrent neural
network (RNN). The training of the set of classifiers can comprise
training the set of classifiers using a dataset with known values
and/or the testing of the set of classifiers comprising testing the
set of classifiers using the dataset with known values. The dataset
can be broken into a first sub-dataset to be used for the training
of the set of classifiers and a second sub-dataset to be used for
the testing of the set of classifiers. The developing of the set of
classifiers can comprise feature selection, and the training of the
set of classifiers comprising normalization of data obtained from a
dataset with known values used to train the set of classifiers. The
using of the set of classifiers to predict the pain level of the OA
patient can comprise using the set of classifiers to predict the
pain level of the OA patient at an Nth visit based on features of
the set of classifiers reported on all visits up to an (N-1)th
visit. The dataset with known values can, for example, the
Osteoarthritis Initiative (OAI) dataset.
[0006] In another embodiment, a method for predicting a pain level
of an OA patient can comprise: developing (e.g., by a processor) a
set of classifiers, the set of classifiers comprising three
classifiers corresponding to a first category, a second category,
and a third category, respectively; training (e.g., by the
processor) the set of classifiers; testing (e.g., by the processor)
the set of classifiers; and using (e.g., by the processor) the set
of classifiers to predict the pain level of the OA patient at a
future visit intended to assess the pain level. The first category
can be that pain has improved for the OA patient since a previous
visit; the second category can be that pain has remained unchanged
for the OA patient since the previous visit; and the third category
can be that pain has worsened for the OA patient since the previous
visit. The developing, training, testing, and using of the set of
classifiers can comprise using an ML technique, such as an eminent
SVM, an RDF, a backpropagation neural network, or an RNN. The
training of the set of classifiers can comprise training the set of
classifiers using a dataset with known values and/or the testing of
the set of classifiers comprising testing the set of classifiers
using the dataset with known values. The dataset can be broken into
a first sub-dataset to be used for the training of the set of
classifiers and a second sub-dataset to be used for the testing of
the set of classifiers. The developing of the set of classifiers
can comprise feature selection, and the training of the set of
classifiers comprising normalization of data obtained from a
dataset with known values used to train the set of classifiers. The
using of the set of classifiers to predict the pain level of the OA
patient can comprise using the set of classifiers to predict the
pain level of the OA patient at an Nth visit based on features of
the set of classifiers reported on all visits up to an (N-1)th
visit. The dataset with known values can, for example, the OAI
dataset.
DETAILED DESCRIPTION
[0007] Embodiments of the subject invention include novel and
advantageous devices and methods for learning and/or predicting the
(self-reported) pain improvement levels of osteoarthritis (OA)
patients (e.g., knee OA patients). A device or apparatus can
include a processor and a machine-readable medium (e.g., a
(non-transitory) computer-readable medium) in operable
communication with the processor and having stored thereon an
algorithm and a unique set of features (see, e.g., Table 2 herein).
The algorithm and/or set of features can be embodied as a set of
instructions stored on the machine-readable medium that, when
executed by the processor, perform steps (including steps of the
algorithm). The algorithm and set of features can enable building
one or more models that learn the (self-reported) pain improvement
levels of OA patients (e.g., knee OA patients).
[0008] In related art devices and methods, patients must physically
visit their medical providers regularly to have an assessment of
their OA status and to report their pain levels at the time of the
visit. This process involves physical tests, imaging, and
demonstrative activities that the patients are asked to perform,
such as chair sits and stands and short distance walks. The pain
level obtained is self-reported by patients, typically using an
outcome scoring system such as the Knee Osteoarthritis Outcome
Score (KOOS) and/or the Western Ontario and McMaster Universities
Osteoarthritis Index (WOMAC). Embodiments of the subject invention
can predict the change in KOOS score for a patient's future visit
using previously measured indicators or features (i.e., from a
previous visit). The algorithm outputs a classification in one of
three categories (improved, unchanged, worsened) corresponding to
the predicted future pain level.
[0009] The algorithm can rely on an ensemble machine learning
approach, such as a Recurrent Neural Network (RNN). The algorithm
can include developing, training, and testing a set of three RNN
classifiers, each corresponding to one of three categories
(improved, unchanged, worsened). In an embodiment, the algorithm
can be trained to predict the reported pain of patients over a span
of time of up to nine years. Three single-class multi-label RNN
classifiers can be elaborated, where a patient is classified into
one of the aforementioned categories. A total of nine labels can be
used, corresponding to the patients' self-reported pain levels
during the nine annual visits. For each label, the classifiers
incorporate the feature values recorded at the time point of the
respective label. Hence, to learn or predict the pain level at the
N.sup.th visit, only features reported up to the (N-1).sup.th visit
are used.
[0010] In an embodiment, in order to produce a complete prediction
of the pain category progress for the OA patients, an ensemble step
of the classifiers corresponding to the three classes (improved,
unchanged, and worsened) can be used. This is necessary because the
three classifiers in each method are independent and only show one
dimension (each) of the prediction result. Final combined
prediction results per algorithm, for an example test, are shown in
Table 6 herein.
[0011] In many embodiments, the algorithm performing the combined
predictions can be summarized as follows: for each of the nine
labels, the algorithm examines the outputs of each of the three
classifiers and takes a weighted vote to determine whether a
patient's pain level has been improved, unchanged, or worsened with
respect to the previous reporting. There are eight possible
scenarios in play, with four distinct cases outlined below: [0012]
1. All three classifiers predict a positive class (that is,
improved, unchanged, and worsened): the algorithm here chooses the
classifier with the highest F1-score as the prediction for that
label. [0013] 2. All three classifiers predict a negative class
(that is, not improved, not unchanged, and not worsened): the
entire data point (i.e., the patient) is marked as a miss. The
algorithm halts for that patient and does not compute any further
predictions for the rest of the labels. [0014] 3. Two of the
classifiers predict a negative class, while the other predicts a
positive result. The classifier with the positive result is chosen
to vote. [0015] 4. Two of the classifiers predict a positive class
while a negative class is predicted by the third: in this case, the
classifier with the highest F1-score is chosen to vote. Here there
are two subcases: first, if the classifier with the highest
F1-score predicts a positive class, its prediction is simply
chosen. Second, if it predicts a negative class, the second-best
performing classifier that predicts a positive class is taken.
[0016] In an embodiment, the publicly available Osteoarthritis
Initiative (OAI) dataset can be used to extract features and to
train and test the models. The datasets can be split into two major
sets (training set and testing set) in order to evaluate the model
over unseen data points. The RNN model of embodiments of the
subject invention was evaluated against the OAI dataset and
compared to three other machine learning models that were also
built. The RNN model achieved an average F1-measure of 0.81 (81%)
on the test set (see Table 6 herein). This embodiment should not be
construed as limiting.
[0017] By utilizing embodiments of the subject invention,
projection of pain outcomes related to OA (e.g., knee OA) can be
improved by an apparatus leveraging existing large databases of
patient data and machine learning techniques. The apparatuses,
devices, and methods of embodiments of the subject invention can
apply machine learning models using, for example, RNNs to predict
the self-reported pain improvement of OA patients (e.g., knee OA
patients).
[0018] Machine learning (ML) approaches to OA diagnosis and pain
predication is relatively unexplored in the related art, but
embodiments of the subject invention can use ML to help distinguish
patients' pain outcome trajectories and improvement given certain
treatments. Embodiments show that projection of pain outcomes
related to knee OA can be improved by leveraging existing large
databases of patient data and ML techniques. The feasibility of
predicting an OA patient's pain improvement or trajectory over nine
years based on a given set of features has been demonstrated.
Several ML techniques and algorithms can be used, including the
eminent Support Vector Machine (SVM), Random Decision Forest (RDF),
and variations of Artificial Neural Network (ANN) algorithms. The
methods leveraged for this task consider the differences in
patients' sex, age, body mass index (BMI), injury factors, and
occupation factors. In addition, the models developed involve
calculation of several features that further include physical and
clinical examination of the patient, including the recorded
physical activity and other self-reported variables.
[0019] Most studies utilizing ML models for OA-related tasks have
focused on image classification. There have been a few attempts to
apply ML to OA-risk identification. These works focused on Logistic
Regression (LR) analysis methods and variants thereof, and they
have been widely used for various prediction and classification
tasks related to OA monitoring and diagnosis such as predicting
outcomes after surgery, risk and pain analysis, as well as
classifying patients as OA patients from others. These statistical
analysis methods proved successful in some cases. However, in most
cases, these are methods that require extensive formal statistical
training, making them far from ideal in a clinical practice. These
methods further proved to be time consuming due to the involvement
many variants and entities that could not be merged to give a clear
result in some cases. LR prediction models involved calculation of
LR equations based on factors such as age, gender and BMI of an
individual. Reports were developed through assessment of clinical
data, physical examination and blood sample for genetic follow
up.
[0020] Researchers have also attempted to build descriptive models
of OA patients based on reported pain. However, most were focused
on the identification of subgroups of patients rather than the
long-term prediction of pain. These studies have examined hip OA,
knee OA, and combinations of both. The methods used for the former
studies have mainly been variations of two-step cluster analysis or
latent class growth modeling. Although some of the studies were
successful in identifying patient subgroups based on pain
trajectories for two to six years, their models were selective in
terms of patient population, limited in time (as most used data
spanning for less than 5 years), and in some cases, ineffective or
ungeneralizable when faced with a new population of patients.
[0021] The process of building, training, and testing 12 ML models
using four ML algorithms will be described below.
Methodology
The Dataset
[0022] The data used in the preparation of the examples were
obtained from the OAI dataset, which is available for public access
at http://www.oai.ucsf.edu. Specific datasets used along with their
respective version numbers are listed in Table 1. The dataset's
cohort consists of an ethnically diverse group of women and men
ages 45 to 79 equally distributed along each age/gender group.
Participants were followed for over nine years for changes in the
clinical status of their respective OA conditions including
worsening, improvement, and onset of symptoms and disabilities.
This was achieved by assessing the patients physically using
traditional methods at the participating clinics in an annual
manner. Information collected included biomarkers, joint symptoms,
general health and function, medication history and inventory, and
physical exam measurements. In total, there were 4,796 patients
enrolled for the baseline visit, which shrunk to 3,444 for the last
recorded annual visit (108.sup.th month).
[0023] The focus of the OAI dataset is on knee OA. Per the OAI
website, "the overall aim of the OAI is to develop a public domain
research resource to facilitate the scientific evaluation of
biomarkers for osteoarthritis as potential surrogate endpoints for
disease onset and progression". Two of the OAI dataset objectives
are to provide data for the purposes of scientific evaluation of
biomarkers for OA, and to support the study of the natural history
of knee OA onset and progression as well as the progression of risk
factors associated with knee OA. Embodiments of the subject
invention also address these two objectives.
Overall Approach
[0024] In many embodiments, the major steps of the approach are as
listed below, starting with the OAI dataset as input and resulting
in an output of labels corresponding to pain categories that can be
assessed: [0025] 1) Data preprocessing--formatting and cleaning;
[0026] 2) Feature selection and representation; [0027] 3) Label
preparation and representation; and [0028] 4) Classification--model
training and testing.
TABLE-US-00001 [0028] TABLE 1 OAI datasets used and corresponding
versions.dagger.. Dataset Release Version Baseline AllClinical
0.2.2 12-month AllClinical 1.2.1 18-month AllClinical 2.2.2
24-month AllClinical 3.2.1 36-month AllClinical 5.2.1 48-month
AllClinical 6.2.2 60-month AllClinical 7.2.1 72-month AllClinical
8.2.1 84-month AllClinical 9.2.1 96-month AllClinical 10.2.2
108-month AllClinical 11.2.1
Data Preprocessing
[0029] Data preprocessing methods can include two steps:
formatting; and cleaning. In formatting, the data obtained from OAI
can be extracted from its original ASCII format into a relational
database to ease programmatic access and manipulation. In cleaning,
1,862 incomplete records (i.e., records with missing data) were
removed. Also, an additional 396 randomly selected records were
removed to ensure a similar gender (53% female and 48% male) and
age distribution to the original dataset. This resulted in 2,538
records (out of 4,796 records originally) used for the remainder of
the process.
Feature Selection and Representation
[0030] The OAI dataset is composed of three major types of data:
categorical variables;
[0031] continuous variables; and imaging-related variables. In many
embodiments, only categorical and continuous variables are
considered. The features selection procedure can eliminate many of
these features from consideration, as described below.
[0032] Feature selection was split into two tasks: first,
individual feature selection, in which 100 features were selected
by hand; and second, automatic feature extraction, which used
statistical methods to further reduce the number of features
used.
[0033] For individual feature selection, a set of 100 features
extracted from 73 variables were identified and a simple
combination of features mentioned. This included demographic
features such as age, gender, race, and ethnicity, in addition to
features regarding the patients' medical history, strength
measures, and physical activity and its respective performance
measures.
[0034] In order to identify the features with the most
discriminative power, two statistical measures were taken at first:
the Fisher coefficient; and the squared Euclidean distance. The
Fisher coefficient represents the ratio between class-variance to
within-class variance, while the squared Euclidean distance is a
widely-used class distance measure. Both methods are commonly used
for the identification of discriminant features. Next, a ranking of
the features was generated based on the methods' criteria.
[0035] Four major data representation techniques were applied to
better represent the features. Although presented in order, the
following methods are independent from each other. First, the
continuous age and BMI, were rescaled to a unified range between 0
and 1. This is especially necessary when variables or features have
widely different scales. For example, the feature age has a real
value between 45 and 79 in OAI while the feature gender is either 0
or 1. If the data is not scaled in this case, the age feature will
overtake the gender feature in terms of importance due to scaling
issues and not because it is more or less significant. Rescaling is
also called normalization. Second, certain features were decomposed
into their independent constituents. For example, the feature
20-meter walk--where the patients were asked to walk 20 meters (m)
while the number of steps and the time taken to complete the task
were recorded--into the features 20-meter walk number of steps and
the feature 20-meter walk time to complete. Third, some features
were aggregated to produce more meaningful features and reduce the
feature space. The features Hip_arthritis, back_arthritis and
hand_arthritis were combined into a single feature called
other_arthritis. Other aggregated features were past medication and
arth_injections. Fourth, all features were binarized; that is, they
were transformed using a binary threshold function where feature
values are either 1 when higher than the threshold and 0 when
lower. Finally, a descriptive list of the features used for the
models developed is shown in Table 2.
Label Preparation and Representation
[0036] The OAI datasets include self-reported pain levels of the
patients. This was collected using the KOOS questionnaire,
including all its subscales. OAI patients were presented this
questionnaire at the baseline visit, as well as at the following
nine annual visits. This data was used as the basis for the pain
label calculation. Starting from the 12.sup.th month visit, labels
per patient per visit were generated as improved, unchanged, and
worsened. These labels were calculated by comparing the
self-reported pain values at the current visit with respect to the
previous visit. The label vectors were binarized with respect the
three categories/labels created to serve the single-class
classifiers. The result is a label vector per record (patient) for
each of the three categories of length 9 representing the follow-up
visits where each value is a 0 or 1 with respect to its category,
that is improved or not, unchanged or not, worsened or not.
Data Preparation for Training and Testing
[0037] The first stage of classification is model training,
followed by a stage of model validation (namely, cross-validation),
and finally a stage of testing on a separate (unseen) set. The
original data was split into two main sets, 80% in a training set
and 20% in testing set (2,030 and 508 records, respectively).
TABLE-US-00002 TABLE 2 List of main features and their decomposed
and sub-features (used in model training and testing) Parent
Feature Sub-features Rank age 45-49 2 50-54 55-59 60-64 65-69 70-74
75-79 gender male 4 female BMI underweight 5 normal overweight
obese class I obese class II obese class III Performance 20-meter
walk: pace (m/sec) 3 measures Single chair stand 3 Repeated chair
stands: able to complete 5 stands 9 Repeated chair stand: pace in
stands/sec 10 physical Kneel 30 minutes or more during single day,
past 30 days 11 activity Get in and out of squatting position 10 or
more times during single day, past 30 days 12 Squat 30 minutes or
more during single day, past 30 days 13 flight of stairs completed,
past 7 days 15 Climb up a total of 10 or more flights of stairs
during single day, past 30 days 14 Lift or move objects weighing 25
pounds or more by hand during single day, past 30 days 17 Medical
RA/other inflammatory arthritis diagnosis 1 History pain medication
use, past 30 days 6 injections for treatment of arthritis, past 6
months 7 Fallen and landed on floor or ground, past 12 months 8
Past pregnancy 16
[0038] Following is a brief description of each of the four ML
models adapted for these examples.
Support Vector Machine
[0039] Support vector machines (SVMs) are supervised ML models
widely used for data analysis regression and classification
applications. One of the advantages of SVMs is that they are
capable of both linear and non-linear classification of data. This
mechanism works in a way whereby data entered is placed
categorically to certain classes that are closely associated. The
model's performance is measured by the systems capability to
predict results using the data presented.
[0040] Three single-class multi-label SVM classifiers were
developed where each patient was classified as one of the following
categories: improved; unchanged; or worsened experienced knee pain.
A total of nine labels were used corresponding to the patients'
self-reported pain levels during the nine annual visits. For each
label, the classifiers incorporated the feature values recorded at
the time point of the respective label. For example, to learn or
predict the pain level at the N.sup.th visit, only features
reported up to the N-1.sup.th visit were used. Because it was not
possible to have a value for the change in pain level at the
baseline visit, three new categories were created to aid in the
prediction of the first annual visit, i.e., the first label
representing the 12.sup.th month visit: high; low; and medium pain.
These acted as values for the previous year pain level feature
discussed earlier.
Random Decision Forest
[0041] Random Decision Forests (RDFs) are ensemble learning methods
and are employed in regression and classification applications.
They operate through the construction of numerous decision trees
during the training stage. The technique outputs the class that
contains the mode of the classes of the collection of collection of
tress. This technique is very influential especially in data mining
applications. A major advantage RDF has over regular decision tress
is that the former method avoids overfitting the training set
unlike the latter. Overfitting is the portrayal of random error and
noise by a statistical model instead of an underlying pattern. This
occurs in complex models where a small number of examples are
presented in relation to the feature space. An overfitted model
results in poor prediction performance and can be sensitive to
minor variations in the training set.
[0042] A similar setup to that of the SVM models was employed here.
Three single-class multi-label RDF classifiers were built based on
the same pain categories discussed earlier. The label space is also
identical to that used in the SVM model.
Backpropagation Neural Network
[0043] Backpropagation Neural Network is another supervised ML
scheme. Backpropagation, alternatively referred to as backward
propagation of errors, is one of the methods used to train ANNs.
The method works hand in hand with various optimization techniques
in a two-phased cycle--propagation and weight update. The technique
works by comparing a newly entered data's output with the existing
data, then performing an error approximation where all the
initially entered data are accredited with the errors equally. The
error can be propagated backwardly to approximate the associated
contribution of error to the original output.
[0044] The same structure was followed for neural network as in the
previous two models, where three single-class multi-label
classifiers were built for the pain categories: improved;
unchanged; and worsened. Some effective variations of this
algorithm can be used. First, an adaptive learning rate was used in
order to avoid oscillation of weights and to improve the
convergence rate at which the network outputs a prediction. Second,
an inertia or momentum variable was integrated, which aids in the
overall performance of the model during training and improves the
learning speed during training. Third, the Nguyen-Widrow layer
initialization function was employed, which is attributed with
drastically decreasing the training time.
Recurrent Neural Network
[0045] Recurrent Neural Networks (RNNs) are yet another type of
ANN. They also utilize the principle of backpropagation of errors
with a slight twist--this is called backpropagation through time.
RNNs are commonly used in speech and text recognition as they are
famed for handling an arbitrary sequence of inputs and outputs.
RNNs have also been used in multiple other applications including
model prediction. The main difference between RNNs and other ANNs
is the internal layer cycling in RNNs, which allows them to perform
well with sequential data. The Long Short Term Memory variant of
RNNs was used. This algorithm alleviates the gradient vanish issue
with RNNs. Finally, the architecture discussed in the previous
models was followed here as well. Three more classifiers were
built, and they were single-class multi-label coupled with the same
structure and processing for the class and label spaces. Therefore,
a total of 12 individual classifiers were developed for this task,
which later were combined in an ensemble fashion to give a single
result for each patient at each of the 9 follow-up visits.
Model Training and Validation
[0046] The SVM classifiers were trained using the RBF kernel
function and a soft margin C of 10,000--a common setup. For the RDF
classifiers, a slightly larger number of parameters to optimize
exists. The max_features parameter was set to the square root of
the total number of features in an individual run, the number of
trees parameter was set arbitrarily to 100, where this is referring
to the number of trees to be built before taking the average of
votes for predictions. Additionally, the min_sample leaf parameter
was set to 50. As for the backpropagation ANNs, all weight
initialization was done using the Nguyen-Widrow layer
initialization function, where the weights are assigned small
random values. The bias parameters were all set to small
non-negative values initially. Further, the adaptive learning rate
was set to 0.01 for the improved-class classifiers, and to 0.1 for
the unchanged and worsened-class classifiers. The momentum value
was set to 0.1 for all three classifiers. The sigmoid function was
used for training. The binary_crossentropy loss function was used,
and Adam's optimization algorithm was followed. The algorithm
showed the optimal parameter values to be 5 for the number of
epochs and 26 for the batch size.
Tools Employed
[0047] Several tools were used to implement the underlying
processes outlined above. For the relational database built during
data preprocessing, PostgreSQL was used. Data analysis was
performed using Java and R, as was training and testing of the
models. Table 3 contains a list of the publicly available packages
and libraries that were used for the training and testing.
Results
Model Validation
[0048] The training set was used for the training as well as for
the validation stages. All 12 individual classifiers were trained
separately on the training set, and later validated using a 10-fold
cross-validation method. All the parameter selection and tuning was
performed with the aid of grid search. Grid search, also called
parameter sweep, is the traditional method used for hyperparameter
optimization that performs exhaustive searching over a predefined
hyperparameter space for a specified learning algorithm.
Cross-validation was performed on the training set as a performance
measure for the hyperparameter optimization and to prevent
overfitting by the SVM and the ANN classifiers. The average
cross-validation results per classifier are presented in Table 4.
All results are presented using the F-measure (also referred to as
F.sub.1-score), which considers the harmonic average of precision
and recall to compute the final score between 0 and 1, where 1 is a
perfect score.
Model Testing
[0049] The 12 classifiers were tested over the testing set only,
which was not introduced to the models previously. All testing was
done in a similar fashion. After the classifiers were presented
with the test data, their generated outputs were compared against
the true label values, which accounted for a hit or miss. The
performance metric used for evaluation is the F.sub.1-score.
Additionally, for comparison, a baseline metric was calculated
using the popular Most Frequent Class technique (MFC). Table 5
shows the testing results for all 12 classifiers per label (the
labels are indexed by visit number, where Visit 1 corresponds to
the 12.sup.th month visit, while the rest of the visits follow
annually). The average column shows the average F.sub.1-scores for
the corresponding classifiers.
TABLE-US-00003 TABLE 3 Software libraries and packages used.
Language Library Modules Java Java-ML.sup.[31] SVM R-
rpart.sup.[32] RDF CRAN e1071.sup.[33] Cross-validation
distribution PARTY.sup.[34] RDF CARET.sup.[35] Cross-validation
kernLab.sup.[36] SVM randomForest.sup.[37] RDF nnet.sup.[38] ANN
(Backprob) rnn.sup.[39] RNN
TABLE-US-00004 TABLE 4 Average cross-validation results during the
training phase. Algorithm Classifier Average (F.sub.1) SVM improved
0.553 unchanged 0.631 worsened 0.627 RDF improved 0.733 unchanged
0.698 worsened 0.826 Backpropagation improved 0.725 ANN unchanged
0.729 worsened 0.819 RNN improved 0.812 unchanged 0.882 worsened
0.856
TABLE-US-00005 TABLE 5 Testing results over the test set per
classifier per visit. Visit Average Algorithm Classifier v01 v02
v03 v04 v05 v06 v07 v08 v09 (F.sub.1) SVM improved 0.4 0.42 0.42
0.42 0.48 0.5 0.5 0.49 0.5 0.458 unchanged 0.5 0.5 0.51 0.51 0.55
0.55 0.56 0.55 0.55 0.531 worsened 0.49 0.52 0.53 0.53 0.55 0.57
0.58 0.58 0.58 0.547 RDF improved 0.58 0.56 0.59 0.59 0.59 0.62
0.62 0.62 0.63 0.6 unchanged 0.64 0.68 0.68 0.68 0.71 0.69 0.67
0.67 0.67 0.676 worsened 0.64 0.64 0.59 0.61 0.61 0.62 0.62 0.62
0.63 0.62 Backpropagation improved 0.4 0.55 0.55 0.55 0.71 0.73
0.73 0.77 0.77 0.64 ANN unchanged 0.5 0.69 0.72 0.76 0.75 0.75 0.79
0.82 0.86 0.737 worsened 0.56 0.66 0.68 0.71 0.71 0.71 0.76 0.81
0.8 0.711 RNN improved 0.61 0.65 0.73 0.73 0.75 0.84 0.85 0.83 0.79
0.753 unchanged 0.76 0.79 0.85 0.85 0.86 0.85 0.87 0.87 0.87 0.841
worsened 0.7 0.83 0.8 0.85 0.89 0.85 0.81 0.83 0.83 0.821
Combination of Models
[0050] In order to produce a complete prediction of the pain
category progress for the OAI patients, a combination step of the
classifiers was added corresponding to the three classes (improved,
unchanged, and worsened). This is important because the three
classifiers in each method are independent and only show one
dimension (each) of the prediction result. The final combined
prediction results per algorithm are shown in Table 6.
[0051] The description of the algorithm performing the combined
predictions is as follows. For each of the 9 labels, the algorithm
examines the outputs of each of the three classifiers and takes a
weighted vote to determine whether a patient's pain level has been
improved, unchanged, or worsened with respect to the previous
reporting. There are 8 possible scenarios in play, with four
distinct ones outlined below: [0052] 1. All three classifiers
predict a positive class (that is, improved, unchanged, and
worsened): the algorithm here chooses the classifier with the
highest F.sub.1-score as the prediction for that label. [0053] 2.
All three classifiers predict a negative class (that is, not
improved, not unchanged, and not worsened): the entire data point
(i.e. the patient) is marked as a miss. The algorithm halts for
that patient and does not compute any further predictions for the
rest of the labels. [0054] 3. Two of the classifiers predict a
negative class, while the other predicts a positive result. The
classifier with the positive result is chosen to vote. [0055] 4.
Two of the classifiers predict a positive class while a negative
class is predicted by the third: in this case, the classifier with
the highest F.sub.1-score is chosen to vote. Here there are two
subcases: first, if the classifier with the highest F.sub.1-score
predicts a positive class, its prediction is simply chosen; and
second, if it predicts a negative class, the second best performing
classifier is taken, which predicts a positive class.
TABLE-US-00006 [0055] TABLE 6 Testing results for combined
classifiers per algorithm. Algorithm F.sub.1 Baseline (MFC) 0.413
SVM 0.502 RDF 0.612 Backprop ANN 0.686 RNN 0.811
[0056] Identifying pain trajectories and predicting pain
improvement of OA patients automatically is of critical
significance (both conceptual and practical) for understanding
pain-related features, as well as the discovery and development of
clinical medicine. Further, this development will aid in
better-informed advice for a personalized treatment plan and on
prognosis given by medical practitioners (trajectories). The
examples focused on knee OA patients in the OAI dataset and
demonstrated the feasibility of using ML to predict the pain
improvement outcomes experienced by OA patients.
[0057] All ML models produced results higher than the baseline
metric. Although the model was the worst performing in terms of the
computation speed, the combined prediction results of the RNN
classifiers proved to perform the best among the rest of the
algorithms with an F.sub.1-score of 0.815, followed by the
backpropagation ANN model at 0.733 F.sub.1. This was also true for
the individual single-class classifiers--the RNN classifiers
outperformed all other models for the three pain classes discussed.
The RNN model performed best due to its distinctive sequential
characteristic, that is, it considers time as a factor in its
prediction. Thus, it is important for ML applications to consider
RNNs when faced with sequential or time-stamped data. The combined
results are close to the averages reported by the individual
classifiers within each algorithm. This explains why the
second--and least desirable--case in the prediction combination
algorithm did not occur often. The cross-validation shows an
approximation of the results reported. Also, the cross-validation
results indicate no model overfitting, which is a common problem
with ML algorithms.
[0058] The classification results of individual labels show an
up-trend for the classification performance over the 9 labels,
where the first visit classification yielded a poorer performance
compared with the next 8 labels. The RDF classifiers are an
exception to this pattern, however. This may be due to its random
nature in selecting an arbitrary set of features to build multiple
decision trees, which repeats at every label producing a similar
performance. The models are improving over time with labels due to
the added feature of previous pain label. In fact, this feature was
selected as the most significant feature by the RDF classifiers
along with related injuries and the BMI values. Moreover, it is
noticeable that the relative performance of the classifiers for
each of the three classes was preserved across the four algorithms
employed. The "unchanged" classifiers performed best, followed by
the "worsened", and then "improved" ones. This is due to the
distribution of patients in the OAI datasets, where more patients
were reporting unchanged levels of pain than the improved and
worsened ones. This further supports the characterization of OA as
a disease of chronic symptoms rather than progressive ones.
[0059] The classifiers built were single-class models, which lead
to an overhead exemplified in the prediction combination algorithm
presented earlier. This can also result in missing data points
entirely due to an ambiguous combined prediction (i.e., not
improved, not unchanged, and not worsened). This can be solved by
transforming the classifiers into multi-class classifiers, which
will reduce the number of models needed to calculate to only a
single classifier per ML method while increasing the amount of
computation time and possibly reducing the performance per model
due to the increased class space. In addition, the models presented
only predict a single time step in the future (i.e., a 12-month
period). This may be improved by identifying and extracting more
discriminant features as well as performing a more extensive and
complex hyperparameter optimization.
[0060] Embodiments of the subject invention capitalize on the
performance of several ML algorithms to highlight the feasibility
of automatic pain improvement prediction of OA patients. This
direction can aid doctors, clinicians, medical students, and even
researchers in disease and associated pain simulation and
prediction.
[0061] The methods and processes described herein can be embodied
as code and/or data. The software code and data described herein
can be stored on one or more machine-readable media (e.g.,
computer-readable media), which may include any device or medium
that can store code and/or data for use by a computer system. When
a computer system and/or processor reads and executes the code
and/or data stored on a computer-readable medium, the computer
system and/or processor performs the methods and processes embodied
as data structures and code stored within the computer-readable
storage medium.
[0062] It should be appreciated by those skilled in the art that
computer-readable media include removable and non-removable
structures/devices that can be used for storage of information,
such as computer-readable instructions, data structures, program
modules, and other data used by a computing system/environment. A
computer-readable medium includes, but is not limited to, volatile
memory such as random access memories (RAM, DRAM, SRAM); and
non-volatile memory such as flash memory, various
read-only-memories (ROM, PROM, EPROM, EEPROM), magnetic and
ferromagnetic/ferroelectric memories (MRAM, FeRAM), and magnetic
and optical storage devices (hard drives, magnetic tape, CDs,
DVDs); network devices; or other media now known or later developed
that are capable of storing computer-readable information/data.
Computer-readable media should not be construed or interpreted to
include any propagating signals. A computer-readable medium of the
subject invention can be, for example, a compact disc (CD), digital
video disc (DVD), flash memory device, volatile memory, or a hard
disk drive (HDD), such as an external HDD or the HDD of a computing
device, though embodiments are not limited thereto. A computing
device can be, for example, a laptop computer, desktop computer,
server, cell phone, or tablet, though embodiments are not limited
thereto.
[0063] It should be understood that the examples and embodiments
described herein are for illustrative purposes only and that
various modifications or changes in light thereof will be suggested
to persons skilled in the art and are to be included within the
spirit and purview of this application.
[0064] All patents, patent applications, provisional applications,
and publications referred to or cited herein (including those in
the "References" section) are incorporated by reference in their
entirety, including all FIGURES and tables, to the extent they are
not inconsistent with the explicit teachings of this
specification.
* * * * *
References